A neural network hardware storage perception compression system and method for a satellite-borne intelligent processing platform

Abstract

The application provides a neural network hardware storage perception compression system and method for a satellite-borne intelligent processing platform, which takes a neural network layer as a basic compression unit, performs quantization and pruning operations on weight bit width and connection structure of the model in parallel, generates a penalty parameter by calculating the difference between the storage size requirement of the compressed network layer and the storage resource of the satellite-borne intelligent processing hardware platform, and then performs constraint on the update of the compression rate and weight during back propagation, so as to realize the perception of the storage resource condition of the target hardware platform during the model compression process; the process changes and updates with the iterative training of the model, and the update process is subjected to hardware perception operation, so as to improve the neural network compression effect and improve the adaptability to different hardware platforms.

Description

Technical Field

[0001] This invention belongs to the technical fields of neural network model compression, spaceborne intelligent processing platforms, and edge deployment of neural networks. Specifically, it relates to a neural network hardware storage-aware compression system and method for spaceborne intelligent processing platforms. Background Technology

[0002] Currently, deploying neural networks for efficient inference on resource-constrained spaceborne processing platforms remains a significant challenge. This is primarily because spaceborne processing platforms, as typical edge processing platforms, face severe limitations in space, size, and power consumption, making them unable to support the massive computational and storage demands of neural network inference. Model compression is an effective method to improve the energy efficiency of neural network inference and reduce the deployment cost at the edge. Pruning and quantization are widely used techniques for effectively reducing model size. However, using a single compression method cannot fully exploit model redundancy. Therefore, combined compression methods using multiple compression techniques have been proposed to obtain lightweight models with extreme value compactness. However, existing combined compression methods currently face two main challenges:

[0003] (1) The joint methods operate relatively independently, failing to leverage the complementary advantages of different methods. Most existing joint compression techniques are staged combinations of different compression methods or tools, meaning that the compression techniques act independently on the neural network model. This fails to achieve the complementary advantages of different compression techniques and is also not conducive to fully exploiting the redundancy of the neural network model, resulting in suboptimal accuracy and model size after compression.

[0004] (2) The compression objective lacks practicality, and the use of only indirect and macro-level compression evaluation metrics is difficult to match with the detailed needs of actual hardware deployment. Existing methods typically use global metrics such as computational complexity, number of weights, and storage space usage to indirectly measure the degree of model compression. This will largely lead to a mismatch between the compressed model and the actual hardware deployment requirements, failing to achieve the goal of improving inference efficiency through actual model compression.

[0005] Therefore, for scenarios involving the deployment of neural networks for inference on resource-constrained spaceborne embedded processing platforms, to address the challenges of existing model compression methods in balancing compression ratio and accuracy, and their poor compatibility with actual hardware deployment conditions, it is necessary to design a collaborative compression method that integrates multiple compression techniques, taking into full account the actual resource limitations of the spaceborne intelligent processing platform. This method aims to reduce model size under extreme hardware resource conditions while achieving an effective trade-off between compression ratio and model accuracy, thereby lowering the hardware cost of deploying neural networks on the spaceborne processing platform. This will support the expansion of applications in satellite-based intelligent image classification, object detection, and language segmentation tasks, providing an effective method and system framework for significantly enhancing the intelligent processing capabilities of satellite edge devices. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a neural network hardware storage and perception compression system and method for spaceborne intelligent processing platforms. This system solves the problems of model deployment difficulties and low inference efficiency caused by limited storage resources when deploying neural networks on spaceborne embedded intelligent processing hardware for intelligent application tasks such as satellite on-orbit remote sensing image classification and target detection.

[0007] This invention is achieved through the following technical solution:

[0008] A neural network hardware storage-aware compression system for spaceborne intelligent processing platforms:

[0009] The compression system consists of three parts: a granularity-configurable pruning module, a precision-configurable quantization module, and a hardware storage scale awareness module.

[0010] The granular configurable pruning module is responsible for pruning operations. It determines which neural network connection structures can be deleted based on the storage requirements of weights and feature maps, and then uses a learnable pruning rate to dynamically adjust the connection structures to reduce the model's storage requirements, eliminate storage bottlenecks, and accurately meet the needs of hardware inference.

[0011] The precision-configurable quantization module performs the quantization operation, which reduces the weights from high precision to low precision to reduce the model's storage requirements and computational complexity; the quantization bit width is determined based on entropy clustering and hardware support, and the quantized weights are output.

[0012] The hardware storage scale awareness module calculates the storage space required for each layer of the neural network, then compares it with the storage resource limitations of the target hardware platform, and calculates a penalty parameter. This penalty parameter is added to the loss function during training to constrain the model's weights and compression ratio, adapting to the storage resources of the hardware platform.

[0013] Furthermore, the granularity configurable pruning module performs a deletion operation on the model connection structure during the weight input process. At the same time, it flexibly configures the pruning granularity according to the calculation mode of the on-board intelligent processor platform when performing neural network inference, and uses configurable pruning granularity to eliminate weight bottlenecks and feature map bottlenecks respectively.

[0014] First, the weights are input into the pruning module to calculate the storage space ratio of the input feature map, output feature map, and weights; then, the storage bottleneck of this layer is determined; finally, a learnable pruning rate is used to dynamically prune the model connection structure.

[0015] The dynamic pruning process can automatically determine the connection structures that need to be deleted or restored as the weights change during model training, and output the pruned weight vector.

[0016] Furthermore, the precision-configurable quantization module is integrated into the weight input of the original network structure and is executed in parallel with the pruning process;

[0017] First, after the weights enter the module, entropy is calculated. Then, based on the entropy and the support of the core processor on the spaceborne intelligent processing platform for inference bit width, a clustering method is used to cluster the weights by entropy to determine different quantization bit widths. Finally, based on the determined quantization bit width, quantization sensing operation is performed on the weights, and the quantized weight vector is output.

[0018] The weight vector output after quantization and pruning is subjected to Hadamard product operation, and the product result is used as the input weight for the final convolution operation.

[0019] Furthermore, the hardware storage scale perception module is used to calculate the storage space consumption of this layer of the neural network and compare it with the actual target spaceborne intelligent processing hardware storage resource constraints to generate corresponding penalty parameters.

[0020] First, the model's input feature map, output feature map, and compressed weights are input into the hardware storage scale awareness module, which combines the quantization accuracy configuration with the actual storage space occupied by computational inference. Then, the penalty parameters for this layer are obtained based on the computational structure and the storage space constraints of the onboard intelligent processing hardware. Finally, the penalty parameters are converted into regularization terms and accumulated into the model's loss function to update and constrain the model's weights and compression ratio during backpropagation.

[0021] A neural network hardware storage-aware compression method for spaceborne intelligent processing platforms:

[0022] The method specifically includes the following steps:

[0023] Step 1, Model Decomposition and Storage Bottleneck Analysis: First, input the original neural network model, and the compression system decomposes the model and calculates the storage requirements of each layer to determine which layers may be storage bottlenecks;

[0024] Step 2, Insert compression module: Insert pruning-quantization co-compression module into the input feature map, output feature map, and weight positions of the neural network layer;

[0025] Step 3: Based on the hardware storage requirements of the input spaceborne intelligent processing platform, configure the compression parameters in the compression module and construct the modified model to be compressed.

[0026] Step 4, Hardware-Aware Model Compression: Perform a model compression process to gradually reduce the model's storage requirements, and iterate training to achieve the desired model accuracy and compression ratio.

[0027] Step 5, Model Reconstruction: Based on the compressed parameter results after training, modify and reconstruct the original model to generate a neural network model suitable for the storage resource limitations of the spaceborne intelligent processing platform.

[0028] Furthermore, in step 3,

[0029] Step 3.1: Input training and testing data and the compressed and modified model; the training data is used for weight updates, the testing data is used to evaluate model performance, and the modified compressed model is used for training.

[0030] Step 3.2: Perform forward inference and calculate the loss function; in each iteration step, calculate the model's output on the training data through forward inference, and then use the loss function to measure the difference between the model output and the true label;

[0031] Step 3.3, Calculate the penalty parameter: In hardware-aware model compression, the penalty parameter is used to constrain the model's weight updates to adapt them to the storage resources of the hardware platform, ensuring that the compressed model remains effective when executed on hardware.

[0032] Step 3.4, Backpropagation; By calculating the gradient of the loss function with respect to the model parameters, backpropagation is performed to adjust the model parameters in the current iteration step to reduce the loss.

[0033] Step 3.5, Update compression parameters and weights: Update compression parameters and weights based on the gradients from backpropagation;

[0034] Step 3.6, determine whether the goal has been achieved: at the end of each iteration step, evaluate the model's performance on the test data and whether it meets the constraints of hardware storage resources; if the predetermined goal is met, end the training; otherwise, continue the iteration.

[0035] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.

[0036] A computer-readable storage medium for storing computer instructions that, when executed by a processor, implement the steps of the above-described method.

[0037] Beneficial effects of the invention

[0038] The present invention has the following beneficial effects:

[0039] 1. By employing a combined approach of pruning and quantization compression methods to compress neural network models, the redundancy of the models can be explored in depth from different perspectives. Compared with a single compression method, the compression ratio is further improved. At the same time, by utilizing the technical complementarity between different compression methods, a good trade-off between compression ratio and accuracy is achieved, thereby improving the deployment and application effect of neural networks on spaceborne intelligent processing platforms.

[0040] 2. The compression process can sense the storage resources of the onboard intelligent processing hardware platform, enabling the compressed model to accurately match the storage consumption during onboard inference. Compared with conventional hardware-aware algorithms, using actual storage scale indicators instead of indirect simulation indicators makes the compressed model more consistent with actual hardware inference needs.

[0041] 3. It adopts a plug-and-play modular software design, which facilitates the construction of compression frameworks and systems based on original network models such as remote sensing target classification, target detection and semantic segmentation, and can quickly generate compressed neural network models suitable for specified spaceborne intelligent processing hardware.

[0042] 4. It can automatically complete the model compression operation for spaceborne intelligent processing hardware platforms without the need for manual fine-tuning. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the compression framework of the neural network hardware storage-aware compression system of the present invention;

[0044] Figure 2 This is a schematic diagram of the method flow of the neural network automated model compression system of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] like Figure 1 , 2 As shown.

[0047] This invention targets spaceborne intelligent embedded processing platforms and, based on pruning and quantization theories, designs a collaborative compression framework for the storage scale of neural network model weights and activation values ​​during inference. It utilizes the storage resources of the target hardware platform as the primary constraint to achieve hardware-aware, automated training-based model compression. Furthermore, this invention employs a plug-and-play modular design, effectively compatible with various neural network architectures. Building the compression framework is convenient and rapid, and it can adapt to spaceborne intelligent processing platforms with different processing units, such as FPGAs, ARMs, and SoPCs. It also supports flexible adjustment of the compression granularity and level.

[0048] This invention combines a theoretical framework with the physical parameters of a practical hardware platform to ensure the system functions correctly in real-world applications. The hardware limitations of this invention include, but are not limited to:

[0049] Storage capacity of the onboard intelligent processing platform: The system needs to consider the available storage space on the hardware platform to ensure that the compressed neural network model can adapt to this limitation.

[0050] Computational performance of the hardware platform: Neural network inference needs to be executed on the processor, so the system needs to ensure that the compressed model can still run efficiently on the platform.

[0051] Storage bandwidth and speed: Hardware parameters related to storage access speed and bandwidth. The system needs to consider the speed at which the model accesses storage to avoid performance lag.

[0052] Hardware-supported quantization bit width: Hardware platforms typically support specific quantization bit widths, such as 8-bit, 16-bit, etc.

[0053] Hardware architecture: The architecture of the hardware platform and the type of processor also affect the system design. Different architectures require different optimization strategies.

[0054] It also calculates communication bandwidth, power and heat dissipation capabilities of the hardware platform, and hardware configurability.

[0055] A neural network hardware storage-aware compression system for spaceborne intelligent processing platforms:

[0056] This invention targets convolutional, fully connected neural network models, constructing a compression system by integrating the model compression process with the training process. The compression system uses neural network layers as the basic compression unit, performing parallel quantization and pruning operations on the model's weight bit width and connection structure. It generates penalty parameters by calculating the difference between the storage requirements of the compressed network layer and the storage resources of the onboard intelligent processing hardware platform. These parameters then constrain the compression ratio and weight updates during backpropagation, enabling the model compression process to be aware of the storage resource situation of the actual target hardware platform. This process continuously changes and updates with the iterative training of the model, and hardware-aware operations are performed on this update process to improve the adaptability of the compressed neural network to different compression ratios.

[0057] The compression system is mainly divided into three parts: a configurable granularity pruning module, a configurable precision quantization module, and a hardware storage scale awareness module; the framework principle block diagram is as follows: Figure 1 As shown.

[0058] The granular configurable pruning module is responsible for pruning operations. It determines which neural network connection structures can be deleted based on the storage requirements of weights and feature maps, and then uses a learnable pruning rate to dynamically adjust the connection structures to reduce the model's storage requirements, eliminate storage bottlenecks, and accurately meet the needs of hardware inference.

[0059] The precision-configurable quantization module is similar to the pruning module. This module performs quantization operations, which reduce the weights from high precision to low precision to reduce the model's storage requirements and computational complexity. It determines the quantization bit width based on entropy clustering and hardware support, and outputs the quantized weights. These weights will be used in convolution operations.

[0060] The hardware storage scale awareness module calculates the storage space required for each layer of the neural network, then compares it with the storage resource limitations of the target hardware platform, and calculates a penalty parameter. This penalty parameter is added to the loss function during training to constrain the model's weights and compression ratio, adapting to the storage resources of the hardware platform.

[0061] The granular configurable pruning module serves as a plug-and-play compression operation module, easily inserted into the original network structure to delete model connections during weight input. Furthermore, the granularity of pruning can be flexibly configured based on the computational mode of the onboard intelligent processor platform during neural network inference. Figure 1 As shown in the figure; since the types of parameters that cause storage bottlenecks in neural network models are different, configurable pruning granularity is used to eliminate weight bottlenecks and feature map bottlenecks respectively;

[0062] First, the weights are input into the pruning module to calculate the storage space ratio of the input feature map, output feature map, and weights; then, the storage bottleneck of this layer is determined; finally, a learnable pruning rate is used to dynamically prune the model connection structure.

[0063] The dynamic pruning process can automatically determine the connection structures that need to be deleted or restored as the weights change during model training, and output the pruned weight vector.

[0064] The precision configurable quantization module is similar to the pruning module, also adopting a plug-and-play modular design, integrated into the weight input end of the original network structure, and executed in parallel with the pruning process.

[0065] First, after the weights enter the module, entropy is calculated. Then, based on the entropy and the support of the core processor on the spaceborne intelligent processing platform for inference bit width, a clustering method is used to cluster the weights by entropy to determine different quantization bit widths. Finally, based on the determined quantization bit width, quantization sensing operation is performed on the weights, and the quantized weight vector is output.

[0066] The weight vector output after quantization and pruning is subjected to Hadamard product operation, and the product result is used as the input weight for the final convolution operation.

[0067] The hardware storage scale perception module is used to calculate the storage space consumption of this layer of the neural network and compare it with the actual target spaceborne intelligent processing hardware storage resource constraints to generate corresponding penalty parameters.

[0068] First, the model's input feature map, output feature map, and compressed weights are input into the hardware storage scale awareness module, which combines the quantization accuracy configuration with the actual storage space occupied by computational inference. Then, the penalty parameters for this layer are obtained based on the computational structure and the storage space constraints of the onboard intelligent processing hardware. Finally, the penalty parameters are converted into regularization terms and accumulated into the model's loss function to update and constrain the model's weights and compression ratio during backpropagation.

[0069] A neural network hardware storage-aware compression method for spaceborne intelligent processing platforms: The method of this invention can be applied to various mainstream neural network models based on convolutional and fully connected structures.

[0070] The method specifically includes the following steps:

[0071] Step 1, Model Decomposition and Storage Bottleneck Analysis: First, input the original neural network model, and the compression system decomposes the model and calculates the storage requirements of each layer to determine which layers may be storage bottlenecks;

[0072] Step 2, Insert compression modules: Insert pruning-quantization co-compression modules into the input feature maps, output feature maps, and weight positions of the neural network layers; the parameters of these modules are configured according to the storage requirements of the hardware platform;

[0073] Step 3: Based on the hardware storage requirements of the input spaceborne intelligent processing platform, configure the compression parameters in the compression module and construct the modified model to be compressed.

[0074] Step 3.1: Input training and testing data and the compressed and modified model; the training data is used for weight updates, the testing data is used to evaluate model performance, and the modified compressed model is used for training.

[0075] Step 3.2: Perform forward inference and calculate the loss function; in each iteration step, calculate the model's output on the training data through forward inference, and then use the loss function to measure the difference between the model output and the true label;

[0076] Step 3.3, Calculate the penalty parameter: In hardware-aware model compression, the penalty parameter is used to constrain the model's weight updates to adapt them to the storage resources of the hardware platform, ensuring that the compressed model remains effective when executed on hardware.

[0077] Step 3.4, Backpropagation; By calculating the gradient of the loss function with respect to the model parameters, backpropagation is performed to adjust the model parameters in the current iteration step to reduce the loss.

[0078] Step 3.5, Update compression parameters and weights: Update compression parameters and weights based on the gradients from backpropagation; these parameter updates are affected by a penalty term to ensure the model adapts to the storage resources of the hardware platform.

[0079] Step 3.6, determine whether the goal has been achieved: at the end of each iteration step, evaluate the model's performance on the test data and whether it meets the constraints of hardware storage resources; if the predetermined goal is met, end the training; otherwise, continue the iteration.

[0080] Step 4, Hardware-Aware Model Compression: Perform a model compression process to gradually reduce the model's storage requirements, and iterate training to achieve the desired model accuracy and compression ratio.

[0081] Step 5, Model Reconstruction: Based on the compressed parameter results after training, modify and reconstruct the original model to generate a neural network model suitable for the storage resource limitations of the spaceborne intelligent processing platform.

[0082] This paper presents an automated model compression system designed for scenarios involving the deployment of neural networks for inference applications on spaceborne intelligent processing platforms. The system builds neural network models based on a compression system. Taking the original neural network model and the storage requirements of the spaceborne intelligent processing hardware platform as input, the system automatically decomposes the model and analyzes storage bottlenecks to generate compression configuration parameters. It then uses a hardware storage scale-aware collaborative compression method to compress model weights. Based on the compression results, the system reconstructs the model, generating a neural network model highly matched to the storage resources of the target hardware platform and suitable for deployment. The principle and flow of the automated neural network model compression system are as follows: Figure 2 As shown.

[0083] like Figure 2 As shown, firstly, the neural network model is input into the compression system. The system decomposes the model according to the network layers and calculates the storage bottleneck of each layer. Secondly, corresponding compression modules are inserted at the weights, input feature maps, and output feature maps of each layer of the model. Then, considering the hardware storage requirements of the onboard intelligent processing platform, the compression parameters in the compression modules are configured to construct the modified model to be compressed. Next, a hardware-aware model compression process is executed, and the expected model accuracy and compression rate are achieved through iterative training. Finally, the original model is modified and reconstructed based on the compression parameter results after training. The resulting compressed model is a neural network model that meets the storage resource constraints of the onboard intelligent processing platform and can be deployed and applied.

[0084] In this embodiment, the application of a convolutional neural network model for on-orbit remote sensing image classification is used as an example to illustrate the specific implementation method, with an FPGA-based onboard intelligent processing platform as the core processor as the deployment target.

[0085] 1. Based on the remote sensing image classification dataset, construct the original convolutional neural network model;

[0086] 2. Using a deep learning framework, a granularity-configurable pruning module, a precision-configurable quantization module, and a hardware storage scale-aware module are constructed respectively.

[0087] 3. Design a neural network structure decomposition function. Using the original neural network model as input, identify the neural network structure through iterative traversal, and read the weights and feature map sizes, calculating the storage ratio.

[0088] 4. Design a compression module insertion function. In the original model, identify the weights and feature maps in the convolutional or fully connected structures through traversal, and insert the corresponding compression module into the corresponding structure;

[0089] 5. Design the compression module configuration function. Configure the compression module according to the model storage scale ratio and the input hardware storage requirements;

[0090] 6. Using the dataset and the modified model as input, design a neural network compression training function, calculate the penalty parameters, and perform backpropagation to update the weights and compression parameters. Output the compression result when the compression ratio and accuracy reach the expected target.

[0091] 7. Based on the above functional modules, according to Figure 2 Build an automated neural network model compression system;

[0092] 8. Input the storage resource requirements of the spaceborne intelligent processing hardware platform, as well as the original convolutional neural network model and remote sensing image dataset, into the automated model compression system to perform compression training operations;

[0093] 9. After the compression training is completed, the compression parameters are obtained, and the model structure is reconstructed based on these parameters to generate the compressed model.

[0094] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.

[0095] A computer-readable storage medium for storing computer instructions that, when executed by a processor, implement the steps of the above-described method.

[0096] The memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDRSDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DRRAM). It should be noted that the memory of the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0097] The foregoing has provided a detailed description of the neural network hardware storage perception compression system and method for spaceborne intelligent processing platforms proposed in this invention, and has elucidated the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A neural network hardware storage-aware compression system for spaceborne intelligent processing platforms, characterized in that: The compression system consists of three parts: a granularity-configurable pruning module, a precision-configurable quantization module, and a hardware storage scale awareness module. The granular configurable pruning module is responsible for pruning operations. It determines which neural network connection structures can be deleted based on the storage requirements of weights and feature maps when the hardware platform deploys neural networks for inference. Then, it uses a learnable pruning rate to dynamically adjust the connection structure, thereby reducing the model's storage requirements, eliminating storage bottlenecks, and accurately meeting the needs of hardware inference. The granularity configurable pruning module performs a deletion operation on the model connection structure during the weight input process. At the same time, it flexibly configures the pruning granularity according to the calculation mode of the on-board intelligent processor platform when performing neural network inference. The configurable pruning granularity is used to eliminate the weight bottleneck and feature map bottleneck respectively. First, the weights are input into the pruning module to calculate the storage space ratio of the input feature map, output feature map, and weights; then, the storage bottleneck of this layer is determined; finally, a learnable pruning rate is used to dynamically prune the model connection structure. The dynamic pruning process can automatically determine the connection structures that need to be deleted or restored as the weights change during model training, and output the pruned weight vector. The precision-configurable quantization module performs the quantization operation, which reduces the weights from high precision to low precision to reduce the model's storage requirements and computational complexity; the quantization bit width is determined based on entropy clustering and hardware support, and the quantized weights are output. The hardware storage scale awareness module calculates the storage space required for each layer of the neural network, then compares it with the storage resource limitations of the target hardware platform, and calculates a penalty parameter. This penalty parameter is added to the loss function during training to constrain the model's weights and compression ratio, adapting to the storage resources of the hardware platform.

2. The compression system according to claim 1, characterized in that: The precision-configurable quantization module is integrated into the weight input of the original network structure and is executed in parallel with the pruning process. First, after the weights enter the module, entropy values ​​are calculated. Then, based on the entropy values ​​and the support of the core processor on the onboard intelligent processing platform for inference bit width, a clustering method is used to cluster the weights by entropy values ​​to determine different quantization bit widths. Finally, based on the determined quantization bit width, the weights are subjected to quantization sensing operation, and the quantized weight vector is output. The weight vector output after quantization and pruning is subjected to Hadamard product operation, and the product result is used as the input weight for the final convolution operation.

3. The compression system according to claim 2, characterized in that: The hardware storage scale perception module is used to calculate the storage space consumption of this layer of the neural network and compare it with the actual target spaceborne intelligent processing hardware storage resource constraints to generate corresponding penalty parameters. First, the model's input feature map, output feature map, and compressed weights are input into the hardware storage scale awareness module, which combines the quantization accuracy configuration to determine the actual storage space occupied by the computational inference. Subsequently, the penalty parameters for this layer are obtained based on the computational structure and the storage space constraints of the onboard intelligent processing hardware. Finally, the penalty parameters are converted into regularization terms and added to the model's loss function to update and constrain the model's weights and compression ratio during backpropagation.

4. A neural network hardware storage-aware compression method for spaceborne intelligent processing platforms, characterized in that: The method is based on the neural network hardware storage-aware compression system for spaceborne intelligent processing platforms as described in any one of claims 1 to 3. The method specifically includes the following steps: Step 1, Model Decomposition and Storage Bottleneck Analysis: First, input the original neural network model, and the compression system decomposes the model and calculates the storage requirements of each layer to determine which layers may be storage bottlenecks. Step 2, Insert compression module: Insert pruning-quantization co-compression module into the input feature map, output feature map and weight positions of the neural network layer; delete model connection structure during weight input, and flexibly configure the pruning granularity according to the computing mode of the on-board intelligent processor platform when performing neural network inference. Use configurable pruning granularity to eliminate weight bottleneck and feature map bottleneck respectively. First, the weights are input into the pruning module to calculate the storage space ratio of the input feature map, output feature map, and weights; then, the storage bottleneck of this layer is determined; finally, a learnable pruning rate is used to dynamically prune the model connection structure. The dynamic pruning process can automatically determine the connection structures that need to be deleted or restored as the weights change during model training, and output the pruned weight vector. Step 3: Based on the hardware storage requirements of the input spaceborne intelligent processing platform, configure the compression parameters in the compression module and construct the modified model to be compressed. Step 4, Hardware-Aware Model Compression: Perform a model compression process to gradually reduce the model's storage requirements, and iterate training to achieve the desired model accuracy and compression ratio. Step 5, Model Reconstruction: Based on the compressed parameter results after training, modify and reconstruct the original model to generate a neural network model suitable for the storage resource limitations of the spaceborne intelligent processing platform.

5. The method according to claim 4, characterized in that: In step 3, Step 3.1: Input training and testing data and the compressed and modified model; the training data is used for weight updates, the testing data is used to evaluate model performance, and the modified compressed model is used for training. Step 3.2, perform forward inference and loss function calculation; In each iteration step, the model's output on the training data is calculated through forward inference, and then the loss function is used to measure the difference between the model output and the true label; Step 3.3, Calculate the penalty parameter: In hardware-aware model compression, the penalty parameter is used to constrain the model's weight updates to adapt them to the storage resources of the hardware platform, ensuring that the compressed model remains effective when executed on hardware; Step 3.4, Backpropagation; By calculating the gradient of the loss function with respect to the model parameters, backpropagation is performed to adjust the model parameters in the current iteration step to reduce the loss; Step 3.5, Update compression parameters and weights: Update compression parameters and weights based on the gradients from backpropagation; Step 3.6, Determine if the objective has been achieved: At the end of each iteration step, evaluate the model's performance on the test data and whether it meets the constraints of hardware storage resources; if the predetermined objective is met, end the training; otherwise, continue the iteration.

6. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method of claim 4 or 5.

7. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method of claim 4 or 5.

Images